Expose callback to overwrite current margin (raw score) during training

[Kim-Jongchan]

I’d like to request support for modifying the current margin (raw score) used in training, before each boosting round.

While DMatrix.set_base_margin() exists, it only sets the initial base margin. During training, the margin is updated via internal prediction buffers which are not exposed or resettable.

Use cases:

Dynamic residual correction

Hybrid models that inject new priors into each boosting step

Time-series modeling where the residual signal evolves

Proposal:

Add a before_iteration callback that allows:

Reading the current margin (model.predict(..., output_margin=True))

Overwriting it via new method (e.g., booster.set_current_margin(...))

or:

Allow flushing internal raw score buffer and reloading from base_margin each round

I’m happy to help test or discuss implementation options. This would make XGBoost much more flexible for advanced workflows.

[trivialfis]

Thank you for raising the issue.

Time-series modeling where the residual signal evolves

Could you please share some references on how this works in practice? I would like to learn more about your use case.